Artificially Intelligent Barbarians
We’re excited about buying and transforming legacy incumbents with AI. What does the AI-led buyout playbook look like?
Yesterday, my colleagues and I published a blog post about AI-enabled services businesses. We’re very excited about their potential: the busted unit economics and scaling challenges that plagued last cycle’s tech-enabled services will be largely solved by LLM-powered agents that can replace certain workers. The piece was well-received by founders and investors alike, and I fielded some questions around how we’re planning to attack the space, so I wanted to share some of my current thinking. Feedback welcomed (particularly from folks who feel differently)!
AI Buyouts: Buy and Build
As we mentioned, we’re particularly excited about a buy-and-build strategy: transform existing services businesses with AI and scale the playbook with M&A. Like PE, but with a huge focus on R&D innovation to drive profitability and improve growth.
Because LLM-driven cost reductions are so compelling, buying and remaking an existing company can make more sense than building something entirely new. Incumbents already own distribution relationships, an increasingly valuable asset as the cost and complexity of building software declines. And these companies have huge corpuses of proprietary data, a critical asset for any business building high-quality AI tools. Of course, managing operationally complex services businesses can be a huge headache – but given how quickly software market structures are changing in the age of AI, the juice is worth the squeeze.
Last September and again in January of this year, I wrote about AI-enabled services as a potentially contrarian opportunity. It was relatively under-discussed then; now, the services opportunity is totally in vogue, with many VC firms openly talking about “services-as-software” and “AI roll-ups” while PE giants like Vista indicate that AI will play a big role in their value-creation process.
This market will soon become crowded by PE firms, search funds, corporate buyers, and even VCs. But we’re still very early, and there’s lots of opportunity. How should investors approach it?
What companies are a good fit?
Target companies where LLMs can “move the needle”
Focus on companies with:
Manual, pen-and-paper workflows: document review, data extraction and entry, simple knowledge retrieval, text-based communication, and copywriting and materials creation
Employees focused on rote process execution, not high-touch advisory and relationship management (the “judgement economy”)
Low margins, with headcount spend representing a significant share of revenue
Markets with labor shortages: the same shortages that made scaling traditional services companies challenging are now a tailwind for AI transformation by creating an incentive to automate
Weak market position: a 4th or 5th player may have far more room for improvement than a market leader and can be acquired for much cheaper
Acquire low-quality companies with white-collar desk employees doing manual, repetitive, rules-based work and avoid companies where LLMs can’t really reduce COGS (like many businesses in the real world). Think insurance TPA and tax prep, not HVAC installation.
Size matters, especially if pursuing a roll-up strategy
Targets need “minimum viable size.” Starting too small is a mistake. AI can’t really have much effect on a two-person company with no fat to trim. Generating a healthy ROI on expensive R&D investments to automate labor requires a sizeable amount of cost-cutting opportunity. And more size also means more distribution and more data: if the goal is to buy distribution and access data that power powerful AI models, buying tiny companies is a wasted effort.
Size also has meaningful implications for investors pursuing a serial acquisition strategy. Companies with a successful transformation playbook can roll up smaller competitors – buying distribution cheaply – and apply the operating system to these businesses. Markets ripe for roll-up activity have companies with owners facing succession challenges (an incentive to sell) and are highly fragmented, with weak competitors and many small companies available to buy. Those who start by buying a large company can benefit from multiple arbitrage: tiny companies with weak competitive positions can be purchased for low multiples but become much more valuable as part of a larger platform with scale advantages. Multiple arbitrage is a key driver of successful roll-up returns; opting for tiny initial acquisitions means foregoing the benefits of financial engineering.
The AI Transformation Playbook
The traditional PE playbook isn’t enough – innovate, don’t just cut
Most buyout firms generate returns via financial engineering (employing leverage, benefitting from multiple arbitrage, improving capital allocation). Some help their portfolio companies with operational restructuring (re-negotiating vendor contracts, eliminating low-ROI marketing and R&D endeavors, transitioning out ineffective personnel and trimming bloat, outsourcing low-level labor to cheaper COL areas) as well. While these are important tools for any transformation, the key component of the AI-enabled services strategy is to fundamentally remake the business with both in-house and third-party AI tools.
Replacing humans with AI agents will create enormous bottom-line ROI. Start by attacking opex like support, sales, and G&A. These functions are the lowest-hanging fruit:
In addition to reducing opex, LLMs can significantly reduce COGS for organizations who sell skilled labor. Purpose-built copilots will first augment the capabilities of existing employees, making them more efficient: consider role-specific tools for CPAs or auditors, for instance. And agents will eventually automate many of these people entirely.
COGS automation may create top-line ROI too. More efficient workers can spend less wasted time on manual processes and more time on building relationships and generating new business. AI will also improve product quality: LLM-generated outputs will be more accurate (and can be delivered more quickly) than human-generated outputs. And because scaling compute involves far less friction than recruiting and retaining new people, services businesses will scale more easily than before. Many companies today are simply supply-constrained; automating employees will remove some of these constraints and allow them to serve many more customers than they do today.
Tightly integrating models into their business creates a huge advantage for AI-first services companies relative to tech-light competitors: when foundation model providers release updates, AI-enabled services companies will benefit from baked-in efficiency improvements.
Internal development or off-the-shelf tools?
Should companies invest in first-party R&D or buy existing solutions from third-party vendors? The calculus will be different for every organizational function. Consider three questions:
What capabilities do third-party products have, and how do those compare to in-house solutions we can build?
How much do these third-party tools cost (effectively operating expenses)?
What is the fully-loaded cost (both up-front and ongoing) of building a robust, feature-complete solution in-house (akin to growth and maintenance capex)?
Many “below the line” functions like S&M, CS, bookkeeping, recruiting, and benefits management are more straightforward than COGS labor involved in product delivery because they’re similar across organizations and don’t require much business-specific context. So many cheap software tools for these functions already exist. For opex, third-party solutions are likely best: building automation tools in-house when cheap and effective solutions exist off-the-shelf may generate poor ROI.
Automating COGS will drive much more impact than automating opex, and it’ll require much more in-house R&D. Relying on third-party agents to automate product delivery – what ultimately defines the business – may not work. To truly produce high-quality, customer-facing outputs, agents will need access to huge amounts of organizational, process, and customer-specific data. Third-party offerings, even those with robust integrations across other organizational datasets, probably aren’t good enough. The best companies in the AI-enabled transformation mold will build their own tools for many functions, maintaining tight control over the quality of the customer experience while still generating tangible ROI.
Striving for 100% automation is a recipe for failure
No technological advancement has ever fully replaced humans: we’re all still here, asking our remote colleagues where they’re calling in from. While the extraordinary pace of AI development may make this wave’s economic and labor market dislocations especially painful in the short term, we shouldn’t expect this time to be different.
For instance, avoid trying to automate relationship-intensive work. People enjoy doing business with other people – especially when that business is complex and requires judgement. AI transformation should target manual, rote tasks that humans hate doing and free them up to do more of the things they excel at: forging relationships, winning business, and exercising discretion around tricky problems.
Transformation should also avoid targeting functions where AI can add significant value only if models perform perfectly and instead focus on automating functions where AI can add significant value even at relatively low performance levels. For instance, companies can safely begin by automating asynchronous, text-based customer support conversations. These automations save significant costs but don’t require extremely high model performance. Conversely, companies should avoid (at least for now) trying to automate outbound sales calls. These automations would create significant value, but the performance requirements are too high: 99% accuracy isn’t good enough, because even a small mistake that revealed the AI would make prospects feel uneasy and betrayed, leading to lower conversion. The risk of automating these functions is still too great.
M&A as a growth strategy
Like all companies, AI-enabled services businesses can grow organically, by buying books of business, or via M&A:
Individual market and customer considerations should drive growth strategy, and many AI-enabled services companies will employ a mix of all three. M&A is particularly exciting, because businesses that demonstrate a clear AI transformation playbook that drives margin expansion can apply that playbook to companies they acquire.
Of course, investors should be cautious. M&A is difficult. Successfully executing a single M&A transaction and integration process presents many challenges: identifying a target, arranging financing, closing a transaction, meshing cultures and processes, managing new teams of people, and implementing change without destroying morale or disrupting the end-customer experience. Doing these things repeatedly as a serial acquirer is even harder – especially when many markets are saturated with PE, search fund, and corporate buyers.
Despite these challenges, it’s possible to successfully execute an AI-forward M&A strategy:
Lean into the competitive advantages of having a compelling, proprietary AI transformation playbook. Because the playbook drives so much operational efficiency, buyers can consistently “overpay” vs. competitors and grow quickly without sacrificing on acquisition ROI. A new, AI-enabled value creation strategy is also an additional returns driver at a time when financial leverage is much more expensive than it was from 2008-2022.
Focus on acquiring LMM or MM businesses (in the roughly $5-25M EBITDA range) that have “minimum viable size,” rather than only tiny SMBs. This limits the amount of M&A execution risk – and reduces competition from both above (large-cap PE) and below (many search funds).
Instill a culture of employee empowerment from the top down. Employees are often resistant to change and will be especially resistant when AI comes for their jobs. Ensuring remaining employees enjoy tangible day-to-day benefits of the more efficient working environment can prevent mass exodus.
Innovate on the 3-5 year (PE) or 7-10 year (VC) holding period: underwriting deals based on the long-term benefits of an AI-forward operating system can give buyers another competitive advantage around entry price.
Data infra is required for AI success – and for measuring the impact of that success
Companies without existing data and analytics instrumentation are a terrible fit for AI transformation – unless their new investors are willing to stand up those systems on Day 1. These systems are a prerequisite for not only effectively leveraging LLMs, but also measuring the impact of AI on operational and financial performance relative to baseline. Models are only as powerful as the data that feed them, and proper analytics are a requirement for truly understanding whether AI transformation is successful. Trying to implement a data-forward business transformation without this infra is a recipe for unclear ROI and ultimately disappointment.
That begs the question: how should investors think about ROI? Consider three impacts:
As a first-order condition of success, margins must improve: is the reduction in labor spend big enough to justify the ongoing cost of third-party tools, up-front investment in first-party solutions, and ongoing development / infra / maintenance cost of code?
Top-line improvements are important, too: can the business serve more customers than before? Are product enhancements enough to justify a higher price? Is the sales team winning more head-to-head business? Is revenue per employee higher?
Minimizing adverse events is critical to successful M&A integration: did customer complaints rise, and was churn elevated? Did NPS suffer? Did employees churn for cultural reasons or frustration with change management?
Existing investment firms should adapt, and new entrants will make their mark
I’m unconvinced that traditional PE and VC firms can successfully execute AI transformations without adaptation – and there’s an exciting opportunity for new entrants to raise dedicated transformation funds that combine the skillsets and ethos of both Wall Street financiers and Silicon Valley speculators.
What will successful transformation funds do well?
Seek buy-in from LPs
While AI transformations look very similar to PE, they have a fundamentally different risk-reward profile than most early-stage startup investments; VCs pursuing the strategy at scale should be thoughtful around portfolio construction and seek alignment from their LPs. It may make sense for VCs to raise dedicated transformation funds.
Ensure economic arrangements reflect reality
Again, PE firms already make investments similar to the AI transformation playbook. But most early-stage deals look very different, so VCs need to innovate on deal structure. Buyout transformations have a narrower dispersion of outcomes than most startup investments – and relatively more of the strategy’s returns are attributable to capital than labor – so VCs should own the majority of the economic upside. Providing 100% of the capital in exchange for only ~15% of the equity makes sense when entrepreneurs are taking enormous technical, execution, or market risk and their labor is responsible for most of the potential value creation. It makes less sense for a buyout and serial M&A strategy.
Aggressively cultivate top-tier AI talent
What VCs lack in financial engineering expertise they make up for in access to top-tier AI talent – the kind of people required to build differentiated in-house AI solutions. Cultivating talent is an area where PE investors will need to do some adaptation of their own. Top technical talent flocks to Silicon Valley or quant trading firms. The buttoned-up culture of most buyout shops isn’t attractive to them. And “operating” talent often takes a backseat – in compensation and career development – to investment professionals. Ensuring talented engineers feel like a priority will be critical to attracting excellent technical talent.
Embrace innovation, but think critically about ROI
Both VCs and PE investors will need to take a slightly different approach to R&D investment. VCs constantly evaluate cutting-edge tech and are comfortable with long-term investments in proprietary tech that reduce near-term FCF: these skills are extraordinarily valuable for AI transformations, which demand a deep understanding of what’s possible and a willingness (and ability) to invest in product development. But many VCs have taken it too far, subsidizing losses that don’t serve a concrete long-term vision. Conversely, traditional financiers are often reticent to make investments that reduce near-term FCF but do have a ruthless eye for ROI. The best investors will be those who marry the investment-forward approach characteristic of VCs with the rigorous focus on tangible ROI typical of traditional buyout firms.
It's so early, and there are so many questions: how much automation potentially really exists? Will third-party tools become powerful enough to drive top-decile returns by themselves? Will PE investors learn tech more quickly than VCs learn financial engineering? We’re eager to meet people excited about these opportunities – if you’re interested in working together, please get in touch!
Thank you to everyone who weighed in on this piece. If you have questions, comments, or feedback, please reach out: andrewziperski [at] gmail [dot] com.
The views expressed herein are the author’s own and are not the views of Craft Ventures Management, LP or its affiliates.
Very insightful!